Research map / Architectures and efficiency / Spiking networks

The Dormant Spiking Neuron: A State-Driven Mechanism for Efficient Spiking Neural Networks

NeurIPS 2026 · Architectures and efficiency › Spiking networks · approach: Training deep spiking networks

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes the Dormant Spiking Neuron (DSN), which permanently gates off postsynaptic neurons whose membrane potential falls below a negative threshold, trained end-to-end with BPTT and surrogate gradients, cutting estimated SNN energy by 40% to over 70%.

Problem

SNNs are efficient only through presynaptic input sparsity, while postsynaptic neurons that are deeply inhibited (negative membrane potential) still incur per-timestep integration and leak costs despite rarely firing.

Approach

A binary dormancy gate per neuron becomes permanently closed once membrane potential drops below a negative threshold Dth, bypassing that neuron's subsequent computation; the gate is integrated into neuronal dynamics and trained end-to-end with surrogate gradients, making sparsity state-driven and orthogonal to input-driven sparsity.

Key results

• On ImageNet, DSN (BPTT) with ResNet-18 reaches 64.44% accuracy at 1.17 mJ, about 73% estimated energy reduction vs. the 4.29 mJ baseline and +1.34% accuracy over the baseline.
• On DVS-Gesture, DSN (BPTT) matches the 96.88% baseline accuracy with 42% less estimated energy.
• Across the threshold sweep, reported energy reductions range from roughly 21% to 73% depending on Dth and dataset; ImageNet ResNet-18 ITO reaches 62.04% at 1.15 mJ.

Surprising findings

• DSN-BPTT on ImageNet improves accuracy over its baseline SNN despite the added dormancy constraint.
• BPTT-trained DSN uses about 34% more GPU memory and 19% lower training throughput than the baseline.

Limitations

• Energy figures are analytical estimates rather than device-level measurements.
• Applicability depends on rate coding or multi-spike redundancy; less suitable for TTFS coding.
• Training costs increase (lower throughput, higher GPU memory), and ImageNet baselines were taken from the literature rather than reproduced.

Experimental setup

Datasets: CIFAR10 CIFAR100 ImageNet DVS-CIFAR10 DVS-Gesture

Benchmarks: CIFAR10 CIFAR100 ImageNet DVS-CIFAR10 DVS-Gesture

Metrics: accuracy estimated energy (mJ) GSOPs GFLOPs neuronal computational utilization spike train accuracy SSIM Jaccard similarity

Baselines: MS-ResNet SD-Transformer AT-LIF DPAP Unstructured pruning (Unstru.) UniSparse SEENN SEW-ResNet Spikformer

Builds on: Spike-Driven Transformer surrogate gradient MS-ResNet SpikingJelly

Closest papers by meaning

• SpikingSSMs: Learning Long Sequences with Sparse and Parallel Spiking State Space Models (AAAI 2025 · similarity 0.90)
• P-SPIKESSM: HARNESSING PROBABILISTIC SPIKING STATE SPACE MODELS FOR LONG-RANGE DEPENDENCY TASKS (ICLR 2025 · similarity 0.89)
• Rethinking the Membrane Dynamics and Optimization Objectives of Spiking Neural Networks (NeurIPS 2024 · similarity 0.89)
• Improving the Sparse Structure Learning of Spiking Neural Networks from the View of Compression Efficiency (ICLR 2025 · similarity 0.89)
• Sign Gradient Descent-based Neuronal Dynamics: ANN-to-SNN Conversion Beyond ReLU Network (ICML 2024 · similarity 0.88)
• NDOT: Neuronal Dynamics-based Online Training for Spiking Neural Networks (ICML 2024 · similarity 0.88)